Matthew Ager

Discriminative and generative machine learning approaches towards robust phoneme classification

Robustness of classification of isolated phoneme segments using discriminative and generative classifiers is investigated for the acoustic waveform and PLP speech representations. The two approaches used are support vector machines (SVMs) and mixtures of probabilistic PCA (MPPCA). While recognition in the PLP domain attains superb accuracy on clean data, it is significantly affected by mismatch between training and test noise levels. Classification in the high-dimensional acoustic waveform domain, on the other hand, is more robust in the presence of additive white Gaussian noise. We also show some results on the effects of custom-designed kernel functions for SVM classification in the acoustic waveform domain.

Combined waveform-cepstral representation for robust speech recognition

High-dimensional acoustic waveform representations are studied as a front-end for noise robust automatic speech recognition using generative methods, in particular Gaussian mixture models and hidden Markov models. The proposed representations are compared with standard cepstral features on phoneme classification and recognition tasks. While lower error rates are achieved using cepstral features at very low noise levels, the acoustic waveform representations are much more robust to noise. A convex combination of acoustic waveforms and cepstral features is then considered and it achieves higher accuracy than either of the individual representations across all noise levels.

High-dimensional linear representations for robust speech recognition

Phoneme classification is investigated in linear feature domains with the aim of improving the robustness to additive noise. Linear feature domains allow for exact noise adaptation and so should result in more accurate classification than representations involving nonlinear processing and dimensionality reduction. We develop a generative framework for phoneme classification using linear features. We first show results for a representation consisting of concatenated frames from the centre of the phoneme, each containing f frames. As no single f is optimal for all phonemes, we further average over models with a range of values of f. Next we improve results by including information from the entire phoneme. In the presence of additive noise, classification in this framework performs better than an analogous PLP classifier, adapted to noise using cepstral mean and variance normalisation, below 18dB SNR.

Problems of Redundancy in Information and Control Systems (RED), 2012 XIII International Symposium on

Automatic speech recognition (ASR) systems are yet to achieve the level of robustness inherent to speech recognition by the human auditory system. The primary goal of this paper is to argue that exploiting the redundancy in speech signals could be the key to solving the problem of the lack of robustness. This view is supported by our recent results on phoneme classification and recognition in the presence of noise which are surveyed in this paper.

Redundancy in speech signals and robustness of automatic speech recognition

Automatic speech recognition (ASR) systems are yet to achieve the level of robustness inherent to speech recognition by the human auditory system. The primary goal of this paper is to argue that exploiting the redundancy in speech signals could be the key to solving the problem of the lack of robustness. This view is supported by our recent results on phoneme classification and recognition in the presence of noise which are surveyed in this paper.